Intelligent medicine management system and method based on multi-modal perception and adaptive decision

Abstract

The application relates to the technical field of intelligent medical equipment, and discloses an intelligent medicine taking management system and method based on multi-modal perception and adaptive decision, which comprises an automatic medicine dispensing terminal with a weighing sensor and an electromechanical locking mechanism, and a wrist device integrating an inertial measurement unit, an acoustic pickup module and an optical perception module. The application establishes a motion timing coherence feature vector by extracting a load step time, a touch action time and a swallowing action time to record a preliminary medicine taking state; an expected concentration change rate is deduced based on a swallowing time to call a pharmacokinetic model, a sampling frequency of the optical module is dynamically adjusted to collect physiological signals; finally, a closed-loop verification state is determined according to a morphological gradient sequence of the physiological signals, and control is executed on the electromechanical locking mechanism. The application can accurately verify a real medicine taking action, optimize hardware power consumption, and effectively prevent a medicine concentration superposition risk caused by delayed medicine taking.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical device technology, specifically to an intelligent medication management system and method based on multimodal perception and adaptive decision-making. Background Technology

[0002] Existing smart medication management devices primarily rely on timed reminders and mechanical switch recording to assist patients in taking medication. These devices typically only monitor the opening status of the pillbox or the removal of medication, and cannot further confirm whether the medication has actually entered the patient's body. Some solutions attempt to introduce a single motion sensor to determine the medication-taking action, but in practical applications, due to the complexity of the action or differences in user habits, single-dimensional data is prone to misjudgment and cannot effectively exclude situations such as simulated medication taking or medication being dropped midway, resulting in insufficient accuracy of the front-end data collection.

[0003] Furthermore, conventional medication management systems often operate on fixed dosing schedules, lacking dynamic consideration of the actual metabolic process of drugs in the body. When a patient delays taking medication, the previously taken drug may still be at a high concentration level or still in the metabolic absorption phase. If the system fails to detect this physiological state and still opens the pillbox at the original time to allow the next dose, the peak release levels of the two doses can easily overlap, potentially leading to pathologically high blood drug concentrations. Most current smart pillboxes fail to integrate multimodal motion monitoring, physiological signal feedback, and pharmacokinetic calculations, and also lack adaptive adjustment mechanisms for dosing time deviations and physical locking interventions at the underlying hardware level. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent medication management system and method based on multimodal perception and adaptive decision-making. This solves the problems of existing technologies that rely solely on single-dimensional mechanical actions or surface data for medication dispensing records, which cannot accurately verify whether the drug has actually entered the digestive tract. Furthermore, it lacks an adaptive adjustment mechanism for dynamic pharmacokinetic changes caused by individual medication time deviations, which can easily lead to the risk of drug concentration accumulation or medication failure.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides an intelligent medication management system based on multimodal perception and adaptive decision-making, comprising: An automated drug dispensing terminal includes a weighing sensor, a dispensing chamber, and an electromechanical locking mechanism for controlling the opening and closing of the dispensing chamber. The wrist-mounted edge computing device is communicatively connected to the automated drug dispensing terminal and integrates an inertial measurement unit, an acoustic pickup module, and an optical hemodynamic sensing module. The system performs the following operations: It establishes a temporal coherence feature vector based on the load step time acquired by the weighing sensor, the contact action time extracted by the inertial measurement unit and the acoustic pickup module, and the swallowing action time; it records the initial medication confirmation status when the components of the feature vector are within a set physiological range; it retrieves a preset pharmacokinetic parameter model based on the swallowing action time to deduce the expected blood drug concentration function and concentration change rate data; it dynamically adjusts the sampling frequency of the optical hemodynamic sensing module and collects physiological signals using the concentration change rate data; it processes the physiological signals to generate a morphological gradient sequence, determines the closed-loop verification pass status based on the morphological gradient sequence, and controls the electromechanical locking mechanism.

[0006] Preferably, the acoustic pickup module includes an electret microphone array and a piezoelectric bone conduction sensor; the steps of extracting the contact action time and the swallowing action time include: intercepting the spatial attitude sequence output by the inertial measurement unit before and after the load step time, performing cross-correlation operation on the spatial attitude sequence and a preset standard spatial motion trajectory sequence to generate a cross-correlation function, and extracting the time node corresponding to the global maximum value of the cross-correlation function as the contact action time; within the static mask effective interval after the contact action time, synchronously acquiring the airborne acoustic wave signal output by the electret microphone array and the mechanical vibration signal output by the piezoelectric bone conduction sensor, calculating the time-domain cross-energy function of the airborne acoustic wave signal and the mechanical vibration signal, and extracting the timestamp corresponding to the time-domain cross-energy function reaching its extreme value as the swallowing action time.

[0007] Preferably, the derivation of the expected blood drug concentration function and the concentration change rate data, as well as the adjustment of the sampling frequency, includes: constructing the expected blood drug concentration function using the extravascular administration model in the pharmacokinetic parameter model; calculating the first-order time derivative of the expected blood drug concentration function to generate the concentration change rate data; extracting the absolute value of the concentration change rate data; converting the absolute value into a dynamic sampling frequency control command using a set mapping function; and overwriting the analog-to-digital conversion sampling rate control register inside the optical hemodynamic sensing module in real time according to the dynamic sampling frequency control command, thereby controlling the actual sampling frequency to be consistent with the value of the dynamic sampling frequency control command.

[0008] Preferably, after determining the closed-loop verification pass status based on the morphological gradient sequence, the intelligent medication management system further performs the following operations: calculates the absolute difference between the maximum and minimum values ​​in the morphological gradient sequence and retrieves the preset basic physiological fluctuation tolerance; when the preliminary medication confirmation status exists and the absolute difference between the maximum and minimum values ​​in the morphological gradient sequence is greater than or equal to the basic physiological fluctuation tolerance, the closed-loop verification pass status is recorded; the actual physical peak time when the morphological gradient sequence reaches its extreme value is extracted, and the numerical difference between the actual physical peak time and the theoretical concentration peak time obtained by deducing the expected blood drug concentration function is calculated; an adaptive update rate constant is introduced based on the numerical difference to iteratively correct the drug absorption rate constant in the pharmacokinetic parameter model.

[0009] Preferably, controlling the electromechanical locking mechanism specifically includes: when the preliminary medication confirmation state exists and the absolute difference of the morphological gradient sequence is less than the basic physiological fluctuation tolerance, determining an abnormal action and controlling the electromechanical locking mechanism to lock the unopened dispensing compartment; when the swallowing action time is later than the planned dosing time node, substituting the iteratively corrected drug absorption rate constant into the pharmacokinetic parameter model to calculate the residual drug load, and performing amplitude superposition calculation of the tail curve containing the decay of the residual drug load and the independent expected blood drug concentration curve generated by the next dosing at the corresponding time node to generate an estimated superimposed peak value; when the estimated superimposed peak value is greater than or equal to the preset safe concentration upper limit data, shifting the simulation start time node corresponding to the next dosing in the positive direction on the time axis according to the set discrete time step and recalculating until the extracted estimated superimposed peak value is less than the safe concentration upper limit data, setting the minimum simulation start time node that meets the conditions as the next safe dosing time node, and controlling the electromechanical locking mechanism to lock to the next safe dosing time node.

[0010] A second aspect of the present invention provides an intelligent medication management method based on multimodal perception and adaptive decision-making, applied to the system described in the first aspect above, comprising: A coherent feature vector of action timing is established based on the load step time obtained from the weighing sensor, the contact action time extracted by the inertial measurement unit and the acoustic pickup module, and the swallowing action time. When the components of the feature vector are within a set physiological range, the initial medication confirmation status is recorded. Based on the swallowing action time, the pharmacokinetic parameter model is retrieved to deduce the expected blood drug concentration function and concentration change rate data. The sampling frequency of the optical hemodynamic sensing module is dynamically adjusted and physiological signals are collected using the concentration change rate data. The physiological signals are processed to generate a morphological gradient sequence. The closed-loop verification pass status is determined based on the morphological gradient sequence, and the electromechanical locking mechanism is controlled.

[0011] Preferably, the extracted contact action time and swallowing action time include: extracting the spatial attitude sequence output by the inertial measurement unit before and after the load step time; performing cross-correlation calculation on the spatial attitude sequence and the standard spatial motion trajectory sequence to generate a cross-correlation function; extracting the time node corresponding to the global maximum value of the cross-correlation function as the contact action time; and within the static mask effective interval after the contact action time, synchronously acquiring the airborne sound wave signal and mechanical vibration signal output by the acoustic pickup module and calculating the time-domain cross-energy function, extracting the timestamp corresponding to the extreme point of the time-domain cross-energy function as the swallowing action time.

[0012] Preferably, dynamically adjusting the sampling frequency of the optical hemodynamic sensing module and acquiring physiological signals using the concentration change rate data includes: constructing the expected blood drug concentration function using the pharmacokinetic parameter model; calculating the first-order time derivative of the expected blood drug concentration function to generate the concentration change rate data; extracting the absolute value of the concentration change rate data; converting the absolute value into a dynamic sampling frequency control command using a set mapping function; and overwriting the analog-to-digital conversion sampling rate control register inside the optical hemodynamic sensing module in real time according to the dynamic sampling frequency control command, controlling the actual sampling frequency to be consistent with the value of the dynamic sampling frequency control command.

[0013] Preferably, after determining the closed-loop verification pass status based on the morphological gradient sequence, the method further includes: extracting the actual physical peak time when the morphological gradient sequence reaches its extreme value; calculating the numerical difference between the actual physical peak time and the theoretical concentration peak time derived from the expected blood drug concentration function; and introducing an adaptive update rate constant based on the numerical difference to iteratively correct the drug absorption rate constant in the pharmacokinetic parameter model.

[0014] Preferably, the control of the electromechanical locking mechanism includes: when the preliminary medication confirmation state exists and the absolute difference between the maximum and minimum values ​​in the morphological gradient sequence is less than the basic physiological fluctuation tolerance, controlling the electromechanical locking mechanism to lock the unopened dispensing compartment; when the actual swallowing action occurs later than the planned dosing time node, substituting the iteratively corrected drug absorption rate constant into the pharmacokinetic parameter model to calculate the residual drug load, and performing amplitude superposition calculation of the tail curve containing the decay of the residual drug load and the independent expected blood drug concentration curve generated by the next dosing at the corresponding time node to generate an estimated superimposed peak value; when the estimated superimposed peak value is greater than or equal to the upper limit of the safe concentration data, shifting the simulated start time node corresponding to the next dosing in the positive direction on the time axis according to the set discrete time step and recalculating until the extracted estimated superimposed peak value is less than the upper limit of the safe concentration data and the next safe dosing time node is obtained, controlling the electromechanical locking mechanism to maintain the locked state until the next safe dosing time node.

[0015] This invention provides an intelligent medication management system and method based on multimodal perception and adaptive decision-making. It has the following beneficial effects: 1. This invention integrates the load step time of the weighing sensor, the spatial attitude of the inertial measurement unit, and the swallowing action time obtained by the acoustic module to establish a coherent feature vector of action timing. This eliminates the limitation of relying solely on a single dimension of mechanical action for judgment, can identify genuine medication behavior, prevents recording errors caused by falsified or interrupted actions, and improves the reliability of data acquisition.

[0016] 2. This invention retrieves pharmacokinetic parameter models based on the swallowing action time, calculates the expected blood drug concentration change rate data, and dynamically overwrites the sampling rate control register of the optical sensing module with this data. This allows the system to increase the sampling frequency in the range of drastic drug concentration changes to capture key physiological signals, and decrease the sampling frequency in the range of moderate concentration changes, thereby controlling the operating power consumption of the hardware module while ensuring monitoring density.

[0017] 3. This invention uses the difference between the actual physical peak time and the theoretical peak time to iteratively correct the drug absorption rate constant, and calculates the residual drug load and estimated superimposed peak in the case of delayed drug administration. When the calculated superimposed peak reaches the safety limit, the system will automatically postpone the next planned drug administration time according to the time step and control the electromechanical locking mechanism to maintain the locked state to prevent the risk of superimposed drug concentration in the body caused by the deviation of drug administration time. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2This is a schematic diagram of the method flow of the present invention; Figure 3 The diagram shows the waveforms for extracting the multimodal action timing features of the present invention, where (a) is a waveform diagram of the tray mass, (b) is a waveform diagram of the normalized cross-correlation coefficient, and (c) is a waveform diagram of the temporal cross-energy. Figure 4 This is a schematic diagram of the relationship between the expected blood drug concentration and the dynamic sampling frequency according to the present invention; Figure 5 This is a schematic diagram of the blood drug concentration extrapolation and electromechanical intervention control curve under the delayed medication scenario of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See attached document Figure 1 , Figure 1 This is a schematic diagram of the system physical topology according to an embodiment of the present invention. The present invention provides an intelligent medication management system based on multimodal perception and adaptive decision-making, comprising: Automated dispensing terminal and wrist-based edge computing device. The automated dispensing terminal and wrist-based edge computing device establish a data connection and maintain clock synchronization via a wireless communication protocol.

[0021] The automated dispensing terminal has an electromechanical control module and a load sensing module. The terminal contains several independent dispensing compartments. Each compartment's outlet is equipped with an electromechanical locking mechanism. The opening and closing of these mechanisms is controlled by a microcontroller outputting a level signal. A miniature load cell is fixedly connected to the bottom tray of the terminal. This load cell outputs real-time tray load mass data via an analog-to-digital converter circuit.

[0022] The wrist-based edge computing device is equipped with a multi-source sensor array. It integrates an inertial measurement unit (IMU) for outputting spatial triaxial acceleration and triaxial angular velocity data. The device also features an acoustic pickup module, comprising an electret microphone array facing the external air environment and a piezoelectric bone conduction sensor facing inwards and conforming to the skin surface. Finally, it includes an optical hemodynamic sensing module. This module contains a photoplethysmography (PPG) analog front-end circuit, supporting dynamic configuration of the LED drive current and the analog-to-digital converter's sampling frequency via read / write register instructions.

[0023] See attached document Figure 2, Figure 2 This is a schematic flowchart of a medication management method according to an embodiment of the present invention. The present invention provides an intelligent medication management method based on multimodal perception and adaptive decision-making, comprising: The system monitors real-time load data of the automated dispensing terminal tray. When a step drop in load is detected and the drop difference matches the set mass of the target drug, the system records the load step time.

[0024] The spatial attitude sequence of the wrist edge computing device before and after the load step time is captured. The spatial attitude sequence is smoothed and filtered, and then cross-correlated with a predefined standard spatial motion trajectory. When the cross-correlation coefficient reaches a set threshold, the time point corresponding to the peak of the cross-correlation function is extracted as the contact action time. The spatial attitude corresponding to this time point is then set as a static mask.

[0025] During the time interval when the static mask is in effect, airborne acoustic signals from an electret microphone array and solid-state mechanical wave signals from a piezoelectric bone conduction sensor are simultaneously acquired. Specific swallowing frequency bands of both signals are extracted using bandpass filters, and the temporal cross-energy function of the two signals is calculated. The first extreme point of the cross-energy function is extracted as the swallowing action time.

[0026] A temporal coherence feature vector is established based on the drug dispensing time, oral contact time, and swallowing time. When the components of this feature vector are within a set physiological range, the system records the initial drug administration confirmation status.

[0027] The system retrieves the pharmacokinetic parameter model for the corresponding drug. Based on the pharmacokinetic parameter model, it derives the expected blood drug concentration function. The first derivative of the expected blood drug concentration function is calculated, generating concentration change rate data.

[0028] The system maps concentration change rate data into hardware sampling control commands. Based on these commands, the sampling frequency of the photoplethysmography (PPG) simulation front-end in the wrist-worn edge computing device is dynamically adjusted. The sampling frequency is increased during periods of high estimated concentration change rate and decreased during periods of low concentration change rate.

[0029] Within the set drug efficacy response time window, the system synchronously acquires photoplethysmography (PPG) signals and triaxial acceleration data. The variance of the acceleration magnitude is calculated to generate a motion intensity index sequence. Using an adaptive filter with the motion intensity index sequence as reference noise, motion decoupling filtering is performed on the PPG signals to obtain purified physiological signals.

[0030] The system performs secondary differentiation processing on the purified physiological signals to generate a secondary differential pulse wave sequence. The absolute amplitudes of the main wave and dicrotic wave in this sequence are extracted, and the reflex index is calculated. By comparing the current reflex index with the initial resting-state reflex index, the morphological gradient data of peripheral vascular resistance are calculated.

[0031] When the morphological gradient data exceeds the baseline physiological fluctuation tolerance within the expected time, the closed-loop validation status of the system is recorded as passed. The actual physical time when the morphological gradient reaches its extreme value is extracted, and the error between this physical time and the theoretical peak concentration time is calculated. This error data is used to update the absorption rate constant in the pharmacokinetic parameter model.

[0032] When the system has a preliminary medication confirmation status but the morphological gradient data has not reached the set threshold, the system determines that the action is abnormal. A locking command is sent to the automated dispensing terminal via the communication link, and the electromechanical locking mechanism locks the unopened dispensing compartment.

[0033] When the actual swallowing action occurs later than the planned time, the system substitutes the updated absorption rate constant into the pharmacokinetic parameter model to calculate the residual drug load in the body. Based on the drug load, the system predicts the next safe dosing time and updates the opening sequence of the corresponding dispensing compartment to the automated dispensing terminal to avoid the superposition of peak blood drug concentrations between adjacent dosing cycles.

[0034] The system microcontroller polls the weighing sensor data of the bottom tray of the automated dispensing terminal at a set sampling frequency. The system calculates the mass difference between adjacent sampling periods. When the mass difference reaches a set step threshold, and the absolute error between this difference and the set reference mass of the target prescription drug is less than the allowable tolerance, the system determines that a drug removal event has occurred. The system extracts the absolute timestamp of the event and records it as the load step time. .

[0035] The system is based on the load step time The system sends a data query command to the wrist-based edge computing device via a wireless communication link. The wrist-based edge computing device then retrieves data from its local memory within the specified time window. Data from the inertial measurement unit within the unit, including The set half-window width is used for truncation. This inertial data contains continuous triaxial acceleration and triaxial angular velocity sequences.

[0036] The system memory contains a pre-stored sequence of standard spatial motion trajectories. This sequence includes three-dimensional kinematic reference features characterizing the continuous physical action of the arm extending, grasping an object, and lifting towards the face. The system calculates the normalized cross-correlation coefficient between the filtered and smoothed spatial attitude sequence and the standard spatial motion trajectory sequence. (Normalized cross-correlation coefficient) The calculation is based on the following formula: ; in, This indicates the current captured spatial attitude sequence at discrete time points. The sampling amplitude; This represents the baseline amplitude of a standard spatial motion trajectory sequence at corresponding discrete time points; The time delay step when performing sliding calculations on two sequences; To capture the total number of data sampling points contained within a time window; To extract the arithmetic mean of the spatial pose sequence; It is the arithmetic mean of the standard spatial motion trajectory sequence.

[0037] The system calculates the normalized cross-correlation coefficient sequence under different time delay steps and extracts the global maximum value. The system compares this global maximum value with a set trajectory matching threshold. When the global maximum value is greater than or equal to the threshold, the system determines that the current arm's spatial motion trajectory conforms to the set medication administration action model. The system extracts the time delay node corresponding to the generation of the global maximum value and, combined with the start time of data extraction, calculates the actual physical time of trajectory peak matching, recording it as the contact point action time. .

[0038] Recording contact action time Subsequently, the wrist-based edge computing device extracts 3D pose Euler angle data for the current time period. The system calculates the angular variance of this 3D pose Euler angle data within a set sliding time window. When the angular variance is lower than a set resting dwell threshold, the system determines that the wrist device is in a relatively spatially static state and sets the spatial pose corresponding to this time period as a static mask. This static mask is used to represent the state in which the patient's arm is physically raised and dwelling in the facial area, and serves as a pre-activation condition for triggering the subsequent acoustic pickup module.

[0039] Once the system determines that the static mask is effective, it triggers the acoustic sensing control logic of the wrist edge computing device. During the time interval during which the static mask remains effective, the wrist edge computing device simultaneously activates the electret microphone array and the piezoelectric bone conduction sensor. The electret microphone array is used to capture sound wave signals propagating through the air medium. The piezoelectric bone conduction sensor, through physical contact, is used to capture low-frequency mechanical vibration signals transmitted along the arm's bones and muscle tissue.

[0040] The system synchronously inputs the acquired acoustic and mechanical vibration signals into their respective bandpass filters. The passband frequency range of the bandpass filters is set to cover the characteristic frequencies generated when a person performs a swallowing action. After filtering, the system removes high-frequency environmental background noise and low-frequency slight body swaying interference from the signal sequence.

[0041] The wrist-worn device is positioned far from the throat, resulting in significant signal attenuation from a single sensor. The system performs cross-modal fusion calculations on the filtered acoustic signal and mechanical vibration signal to generate a time-domain cross-energy function. This calculation is based on the physical property that airborne sound waves and solid-state vibrations overlap concurrently on the time axis during swallowing. (Time-domain cross-energy function) The calculation process is as follows: ; in, Indicates the current calculation time point; This indicates the set length of the sliding integral time window; This indicates that the electret microphone array is in the integral variable The amplitude of the output acoustic signal; This indicates the integral variable of the piezoelectric bone conduction sensor. The amplitude of the mechanical vibration signal output at the point; This represents the impulse response function of the set bandpass filter; This represents the convolution operator; This indicates taking the absolute value of the signal.

[0042] The system evaluates the output of the time-domain cross-energy function sequentially along the time axis. Since the cross-energy algorithm performs the product integral operation of the absolute values ​​of the two signals, the calculated cross-energy result will show a significant jump if and only if the acoustic signal and the mechanical vibration signal synchronously reach their energy peaks within the same time window.

[0043] The system continuously compares the output cross-energy values ​​with a set energy threshold. When the cross-energy value rises and crosses the energy threshold, the system determines that a valid swallowing event has been captured. The system then performs an extreme value search within the data range triggered by this event, extracting the physical timestamp corresponding to when the cross-energy function reaches a local maximum. The system records this physical timestamp as the swallowing event time. After acquiring the swallowing time, the system shuts off the power to the electret microphone array and the piezoelectric bone conduction sensor.

[0044] The system extracts the recorded load step time, contact action time, and swallowing action time. The system performs time difference calculations on the above three time points to calculate the first time interval from drug extraction to the wrist approaching the face, and the second time interval from the wrist approaching the face to the swallowing action.

[0045] The system constructs a temporal coherence feature vector of actions based on the first time interval and the second time interval. This feature vector... The construction is based on the following formula: ; in, Indicates the load step time; Indicates the contact action time; Indicates the time of the swallowing action.

[0046] The system memory pre-writes an action time threshold distribution model. This model consists of the time span boundaries of real human beings completing continuous drug administration actions, obtained through statistical analysis. The system extracts each time difference component from the action temporal coherence feature vector and compares it numerically with the upper and lower threshold limits set by the action time threshold distribution model.

[0047] When either the first or second time interval is negative, it indicates that the timing of the physical actions captured by the device violates causal logic. When the time difference component exceeds the set threshold upper limit, it indicates that there is a non-physiological pause delay between consecutive physical actions. In any of the above time anomalies, the system determines that the current action sequence is invalid and records the physical behavior chain as broken.

[0048] When all components of the action-time coherence feature vector are within the set physiological range, the system determines that the acquired spatial trajectory sequence and acoustic vibration characteristics conform to a real and coherent dynamic medication administration pattern. The system writes a verification flag into the microcontroller's storage space to record the initial medication confirmation status. This confirmation status serves as a prerequisite for lifting hardware power consumption limits and initiating subsequent biological metabolic data sampling.

[0049] After recording the initial medication confirmation status, the system triggers the pharmacokinetic deduction logic of the wrist-based edge computing device. Based on the target prescription drug identification code set in the automated dispensing terminal, the system retrieves the corresponding pharmacokinetic parameter model. The pharmacokinetic parameter model contains physiological constants characterizing the process by which a specific drug enters the human circulatory system. These constants include the drug's absorption rate constant, elimination rate constant, bioavailability, and apparent volume of distribution.

[0050] The system uses the swallowing time as the physical starting point for drug entry into the gastrointestinal tract and the commencement of absorption and metabolism. The system utilizes a one-chamber extravascular drug delivery model to construct a continuous function of the expected blood drug concentration over time. This expected blood drug concentration function... The calculation is based on the following formula: ; in, Represents the current system absolute time; This indicates the recorded swallowing time; Indicates the bioavailability of the extracted drug; This indicates the physical dosage for a single administration of medication; This represents the drug absorption rate constant for that patient stored in the database; This represents the elimination rate constant of the drug; It represents the apparent volume distribution.

[0051] Conventional wearable devices typically employ a fixed low-frequency sampling mechanism to maintain continuous battery life. However, this mechanism cannot fully capture the transient physiological waveform deformations induced by pharmacological effects. To establish a cross-domain scheduling mapping relationship between underlying hardware sampling resources and human physiological changes, the system extracts dynamic intervals where physiological parameters show significant changes. The system calculates the first-order time derivative of the expected blood drug concentration function with respect to the current system time, generating concentration change rate data. This concentration change rate data... The calculation process is as follows: ; The concentration change rate data output by the system reflects the gradient of drug concentration increase or decrease in the blood per unit time. A large absolute value of the concentration change rate indicates that the drug is in a rapid entry into the bloodstream or rapid metabolism phase, during which the hemodynamic characteristics of peripheral blood vessels change most dramatically, requiring the system to have high biosignal capture resolution. When the absolute value of the concentration change rate approaches zero, it indicates that the blood drug concentration has reached a steady-state peak or entered a plateau clearance phase. The system writes the real-time calculated concentration change rate data into a designated buffer area of ​​the microcontroller, serving as a direct control variable for subsequent interventions in the underlying working cycle of the photoelectric sensing module.

[0052] The system extracts concentration change rate data stored in the microcontroller's buffer. It performs absolute value calculations on the concentration change rate data to obtain a non-directional scalar representing the severity of blood drug concentration fluctuations. Using a predefined mapping function, the system converts this absolute value data into sampling control commands adapted to the underlying hardware.

[0053] The mapping transformation process is calculated according to the following formula: ; in, This indicates the dynamic sampling frequency control command output by the system; This represents the absolute value of the concentration change rate data; This represents the set baseline sampling constant, used to maintain the device's base power consumption during the drug-free response period; This represents the set proportional mapping gain coefficient; This indicates the lowest effective sampling frequency physically supported by the analog front-end circuitry in a wrist-based edge computing device; This indicates the highest effective sampling frequency physically supported by the analog front-end circuit; This represents the operation of retrieving the maximum value from the set; This indicates the operation of finding the minimum value in a set.

[0054] The system microcontroller acquires the dynamic sampling frequency control command from the computational output. The microcontroller encodes the dynamic sampling frequency control command into a register write operation data frame supported by the peripheral device communication protocol. The microcontroller then sends this data frame to the photoplethysmography (PPG) analog front-end circuit within the optical hemodynamic sensing module via the internal integrated circuit bus or serial peripheral interface bus.

[0055] The logic control unit of the analog front-end circuit parses the received data frames and performs real-time overwrite operations on its internal registers. Specifically, the real-time overwrite operation includes writing to the analog-to-digital converter sampling rate control register and the LED pulse timing register. By changing the value in the analog-to-digital converter sampling rate control register, the system adjusts the frequency division factor of the internal system clock of the analog front-end, ensuring that the actual output data rate of the analog-to-digital converter matches the value of the dynamic sampling frequency control command.

[0056] The system synchronously overwrites the LED pulse timing register to increase the LED drive pulse repetition frequency. This control logic ensures that the refresh frequency of the optical transmitter is physically synchronized with the sampling frequency of the analog-to-digital converter.

[0057] When the expected blood drug concentration experiences a significant increase or rapid elimination, the absolute value of the rate of concentration change increases. The system outputs a high-value sampling frequency control command through the aforementioned mapping mechanism. The analog front-end circuit increases the analog-to-digital conversion sampling frequency according to this command. The system captures a high-temporal-resolution photoplethysmography (PPG) waveform sequence at the hardware level. When the drug effect enters a stable metabolic phase, the absolute value of the rate of concentration change approaches zero. The system controls the sampling frequency of the analog front-end circuit to fall back to the basic sampling baseline constant, reducing the workload of the microcontroller and RF transmission module.

[0058] Within the set drug efficacy response time window, the system synchronously acquires the raw pulse wave sequence output from the photoplethysmography (PPG) simulation front end and the triaxial acceleration sequence output from the inertial measurement unit at a dynamically adjusted sampling frequency. The raw pulse wave sequence is superimposed with baseline drift and waveform distortion caused by changes in the sensor-skin contact state due to human physical activity.

[0059] The system calculates the Euclidean magnitude of the triaxial acceleration sequence at each discrete time point. The extracted Euclidean magnitude sequence is then input into a predefined sliding calculation window to calculate the local variance of the data within that window. By continuously calculating the local variance of each sliding calculation window, the system generates a motion intensity index sequence. This sequence quantitatively characterizes the degree of mechanical disturbance caused by physical activity to the microcirculatory blood flow volume of the body surface.

[0060] The system constructs a normalized least mean square adaptive filter within the microcontroller of the wrist-based edge computing device. The system sets the original pulse wave sequence containing physical interference as the desired signal input to the adaptive filter, and sets the motion intensity index sequence for the corresponding time period as the reference noise input. The adaptive filter iteratively updates its internal tap weight vector to ensure that the output signal generated after filtering the reference noise approximates the motion artifact components in the original pulse wave sequence.

[0061] The system performs parameter update calculations for the normalized least mean square adaptive filter point-by-point according to the discrete time series. The iterative process of the filter tap weight vector is based on the following formula: ; in, Indicates at discrete time nodes The filter tap weight vector; Indicates at time node Updated tap weight vector; This represents the set filter iteration step size constant; Indicates at time node The extracted reference noise input vector is composed of motion intensity index data of the current time node and historical lag time nodes; The squared Euclidean norm of the reference noise input vector is represented by ; This represents a very small positive number, used to prevent the divisor term in the formula from having a zero value; Indicates at time node Calculate the generated filter error signal.

[0062] The system calculates the difference between the desired signal and the current adaptive filter output signal, and generates the filtering error signal corresponding to that time point. As the time series progresses and iterative calculations are executed, the filter tap weight vector converges, and the filter's output signal approximates the actual motion interference waveform. The system extracts the filter error signal under stable convergence state as the decoupled purified physiological sequence.

[0063] The purified physiological sequence eliminated non-pharmacological signal mutations caused by arm waving or limb displacement, while retaining the pulse waveform characteristics corresponding to the peripheral vasomotor state induced by drug action.

[0064] The system acquires the purified physiological sequence output by the adaptive filter. The system then uses a second-order central difference algorithm to perform discrete-time numerical differentiation on the purified physiological sequence, generating a second-order differential pulse wave sequence.

[0065] The system employs an extreme value search algorithm to locate the peaks of the quadratic differential pulse wave sequence on a per-cardiac cycle basis. The system extracts the absolute amplitude of the main peak of the early systolic positive wave, which characterizes the ventricular ejection state, from a single cardiac cycle sequence. Simultaneously, the system extracts the absolute amplitude of the dicrotic peak of the early diastolic positive wave, which characterizes the peripheral arterial reflection wave state, within the same cardiac cycle.

[0066] The system calculates the reflex index based on the absolute amplitudes of the dominant peak and the dicrotic peak. This reflex index data is used to quantitatively characterize the physiological state of peripheral vascular resistance at the current moment. The calculation is based on the following formula: ; in, Indicates the current system time; This represents the absolute amplitude of the diabetic peak within the cardiac cycle corresponding to the current discrete time point; It represents the absolute amplitude of the main peak of the early systolic wave within the same cardiac cycle.

[0067] The system retrieves resting-state photoplethysmography (PPG) data recorded by the wrist edge computing device before the initial drug administration action. Following the same processing steps described above, the system extracts the resting-state reflectance index and sets it as the baseline reflectance index constant. The system calculates the numerical difference between the reflectance index at the current time point and the baseline reflectance index constant, generating morphological gradient data. This calculation is performed according to the following formula: ; in, Indicates the system time The calculated morphological gradient data of peripheral vascular resistance; This represents the baseline reflectance index constant recorded by the system.

[0068] The system continuously performs the above calculations within the sampling period of the pharmacodynamic response time window, generating a morphological gradient sequence. This sequence excludes the baseline effects of individual baseline heart rate and blood pressure fluctuations, retaining only the pulse waveform deformation component induced after the drug is absorbed into the bloodstream. The system iterates through and compares the morphological gradient sequence within the pharmacodynamic response time window, extracting data points where the sequence values ​​reach their maximum or minimum extreme values. The system extracts the system's absolute timestamp corresponding to this extreme data point and records it as the actual physical peak time. This actual physical peak time characterizes the true moment when the drug exerts its maximum pharmacological intervention effect on the human peripheral cardiovascular system. The system writes the actual physical peak time into a shared storage area as a physiological basis for subsequent validation of the dosing loop and updating of the pharmacokinetic model.

[0069] At the end of the set drug efficacy response time window, the system performs a comprehensive status determination on the execution results of the multi-source data link. The system extracts the preliminary medication confirmation status flag recorded in the system memory and simultaneously extracts the continuous morphological gradient sequence generated within that time window. The system calculates the absolute difference between the maximum and minimum values ​​in this morphological gradient sequence.

[0070] The system compares the calculated absolute difference with the set baseline physiological fluctuation tolerance. When the system has a preliminary medication confirmation status and the absolute difference of the morphological gradient sequence is less than the baseline physiological fluctuation tolerance, the system determines that the peripheral cardiovascular system has not produced the expected pharmacological response. This data indicates a logical causal break between the physical action trajectory and the body's drug metabolism. Based on this contradictory state, the system determines that the user has violated the rule of not taking medication, and the state machine transitions from the monitoring operation state to the abnormal lockout state.

[0071] In an abnormal lockout state, the microcontroller of the wrist edge computing device constructs a data frame containing an abnormal error code and the current timestamp. The wrist edge computing device then transmits this data frame to the automated drug dispensing terminal via a wireless communication link.

[0072] The microcontroller inside the automated dispensing terminal receives and parses the data frame via a communication interface. Upon detecting an error code, the microcontroller immediately triggers its internal high-priority interrupt service routine. Within the interrupt service routine, the microcontroller controls the level states of designated general-purpose input / output pins connected to it.

[0073] The microcontroller pulls the pin level of the corresponding dispensing compartment control channel from the driven state to the cut-off state. This pin level change causes the relay drive circuit connected to the dispensing compartment door to disconnect. The solenoid valve push rods located at the outlets of each undispensed compartment are driven by the built-in springs to return to the mechanically extended position due to the de-energization of the coils. This mechanical extension action forcibly and rigidly locks the physical door of the dispensing compartment, cutting off the physical path for subsequent drug acquisition.

[0074] While performing the hardware locking action, the system generates an alarm log containing the device number and the time of the violation.

[0075] The system compares the calculated difference between the morphological gradient sequences and the set baseline physiological fluctuation tolerance. When the absolute difference in the morphological gradients is greater than or equal to the baseline physiological fluctuation tolerance, the system determines that the peripheral cardiovascular system has produced a pharmacological response that matches the expectation. The system control state machine changes to a closed-loop verification passed state and records that the medication administration was effective.

[0076] The system extracts the actual physical peak time recorded in the previous stage. Simultaneously, the system retrieves the concentration change rate data calculated and output by the microcontroller. Within the corresponding drug efficacy time window of this data sequence, the system searches for system time nodes where the concentration change rate approaches zero or where the positive and negative polarities reverse. The system extracts this time node and records it as the theoretical concentration peak time. This theoretical concentration peak time characterizes the moment when the drug efficacy reaches its maximum value as derived under the initial general model.

[0077] The system calculates the numerical difference between the actual physical peak time and the theoretical peak time. This difference reflects the deviation between the patient's actual gastrointestinal metabolic absorption capacity and the general pharmacokinetic model. The system incorporates this deviation into an iterative algorithm to perform numerical corrections on the physiological constants recorded in the system's database. Through iterative corrections, the system ensures that the underlying pharmacological model aligns with the patient's actual metabolic patterns.

[0078] The iterative correction process for this parameter is based on the following formula: ; in, This represents the drug absorption rate constant updated after this closed-loop calculation; This represents the drug absorption rate constant for this patient currently stored in the system database; This represents the adaptive update rate constant set by the system, used to control the step size limit for convergence of a single iteration; This indicates the actual physical peak time extracted by the system; This indicates the calculated peak time of the theoretical concentration.

[0079] The system microcontroller writes the updated drug absorption rate constant into the database system, overwriting the original constant record. The system will directly call the updated constant data the next time it executes medication scheduling and dynamic sampling frequency simulation. By performing the above comparison and calculation process after each effective medication administration, the physiological model called at the system's underlying level gradually converges from a population-wide model to an individual-customized model.

[0080] The system retrieves the planned dosing time sequence from the memory. The system extracts the actual swallowing time recorded in previous periods. The system calculates the time difference between the actual swallowing time and the planned dosing time for the current cycle. When the actual swallowing time is later than the planned dosing time and the time difference exceeds the set delay tolerance threshold, the system determines that a delayed medication has occurred in the current cycle and triggers the anti-drug effect superposition logic.

[0081] The system calls a closed-loop verification mechanism to update the drug absorption rate constant and the corresponding drug elimination rate constant. The system uses the actual swallowing time as the starting point and substitutes the updated constants into the pharmacokinetic parameter model. The system calculates the residual drug load in the patient's body when the system time advances to the next planned dosing time. This residual drug load is calculated by discretizing and numerically integrating the single physical dose combined with an exponential decay factor and a metabolic elimination factor based on the actual time difference.

[0082] The system obtains the physical dosage set for the next planned dosing cycle. It then uses a pharmacokinetic parameter model to generate an independent expected plasma concentration curve for the next dose. Under a unified time axis, the system performs amplitude superposition calculations on the tail-end curve (containing the residual drug load and its continuous metabolic decay) and the independent expected plasma concentration curve generated for the next dose, calculating the amplitude at corresponding time points. The system performs an extreme value search on the superimposed composite plasma concentration curve and extracts the global maximum value as the estimated superposition peak value.

[0083] The system reads the pre-written safe concentration limit data for the corresponding drug from the database. The system compares the estimated cumulative peak value with the safe concentration limit data. When the estimated cumulative peak value is greater than or equal to the safe concentration limit data, the system determines that opening the dispensing warehouse at the originally scheduled next dosing time poses a risk of pathologically increasing blood drug concentration.

[0084] The system initiates a dynamic postponement algorithm. Using the upper limit of the safe concentration as the target constraint, the system shifts the simulated start time node corresponding to the next dosing dose forward on the time axis according to the set discrete time step. After each shift, the system recalculates the superimposed composite blood drug concentration curve and extracts a new estimated superimposed peak value. The system continuously executes this shift and recalculation cycle until the extracted estimated superimposed peak value is less than the set upper limit of the safe concentration. The system then extracts the smallest simulated start time node that satisfies the constraint and sets it as the next safe dosing time node.

[0085] The system generates a time-updated data frame based on the next safe drug administration time point. The wrist-worn edge computing device uses a wireless communication protocol to send the time-updated data frame to the automated drug dispensing terminal. The microcontroller of the automated drug dispensing terminal receives and parses the data frame, extracting the updated target timestamp.

[0086] The microcontroller uses the extracted target timestamp to overwrite the alarm comparison register in the internal real-time clock module associated with the next target dispensing bay opening task. The microcontroller maintains the cut-off level state of the corresponding electromechanical locking mechanism. The automated dispensing terminal rigidly locks the target dispensing bay until the system physical time reaches the next safe dosing time node, at which point the microcontroller outputs a drive level to trigger the relay to release the solenoid valve push rod.

[0087] Furthermore, it should be understood that the implementation of some underlying drivers and basic algorithms involved in the embodiments of this invention includes, but is not limited to: filtering preprocessing for inertial data, configuration of the digital transfer function of the bandpass filter, configuration of the order of the adaptive filter and microcontroller fixed-point conversion, quadratic discrete difference calculation of digital signals, log uploading based on the standard transmission control protocol, erase / write and verification logic of the underlying non-volatile memory, and timer and interrupt vector configuration of the real-time clock module. The above operations are all conventional techniques that can be directly implemented by those skilled in the art based on microcontroller datasheets, conventional digital signal processing theory, and known communication protocols. To highlight the core innovation of this invention and make the specification more concise and clear, this document does not exhaustively describe the above-mentioned known details, but this does not constitute a limitation on the completeness and implementability of the technical solution of this invention.

[0088] Specific application examples: Target patient: A 75-year-old male with primary hypertension and mild cognitive decline (prone to forgetting or taking the wrong medication).

[0089] Prescription medication: Nifedipine controlled-release tablets (a potent calcium channel blocker that dilates peripheral blood vessels and reduces peripheral vascular resistance, perfectly suited to the monitoring logic of this system based on the reflectance index and morphological gradient of photoplethysmography). Single dose: 30 mg.

[0090] Dosing schedule: Twice daily (BID), scheduled at 08:00 and 20:00. Upper limit of safe blood concentration (UPC) The concentration was set to 80 ng / mL.

[0091] 8:00: The system reminded the patient to take medication, but the patient did not respond.

[0092] 10:00: The patient finally walks towards the equipment.

[0093] Load step ( At 10:00:10, the system detected a step drop of 30.5 mg in the tray (within tolerance), and extraction was performed. .

[0094] Contact action ( Wrist IMU cut-off The data before and after the event were cross-correlated with the standard trajectory. At 10:00:14, the cross-correlation coefficient reached 0.92 (exceeding the threshold), and the contact action was recorded. And apply the static mask.

[0095] Swallowing action ( Within the static mask, acoustic and vibration sensors are activated synchronously. At 10:00:16, a significant local extremum appears in the time-domain cross-energy function, and the swallowing time is recorded. .

[0096] Timing determination: First time interval ( ), second time interval ( All results were within the physiological range, and the initial medication confirmation status was recorded.

[0097] Dynamic sampling trigger: The system is fed in the initial universal pharmacokinetic parameters of nifedipine (absorption rate). The simulation revealed a high rate of change in blood drug concentration within 1-3 hours after administration. The system then sent a command to the simulation front end to dynamically increase the PPG sampling rate from the basic 20Hz to 100Hz in order to capture high-quality waveforms.

[0098] Morphological gradient extraction: The system continuously extracts the patient's PPG signal, and calculates the peripheral reflex index after motion decoupling filtering and second-order differentiation. It was found that 1.5 hours after medication, the reflex index decreased significantly, and the absolute difference exceeded the baseline physiological fluctuation tolerance. The closed-loop verification was passed, confirming that the patient not only put the medication in their mouth, but that the medication had actually entered the bloodstream and taken effect, ruling out the possibility of faking medication.

[0099] Model Correction: The theoretical peak time is 11:30. However, the actual extreme point of the reflectance index extracted by the system occurs at 12:00 (i.e., the actual physical peak time). Due to the weakened gastrointestinal absorption in the elderly, the system adjusts the error formula accordingly. The patient's specific absorption rate constant was lowered, achieving individualized convergence.

[0100] Scheduling simulation: Since the morning medication was delayed by 2 hours (taken at 10:00), if the evening medication is taken at 20:00 as originally planned, the system uses the updated constant integral to calculate and finds that the residual load in the body is high at 20:00. After adding the evening dose, the peak value will reach 85ng / mL (exceeding the safe limit of 80ng / mL, which can easily cause hypotension and syncope).

[0101] Dynamic postponement: The system performs a translation optimization, ultimately rigidly postponing the opening time of the next dispensing bay to 22:30. The electromechanical locking mechanism of the dispensing terminal maintains a forced rigid lock at 20:00, cutting off the possibility of cumulative poisoning from the physical source.

[0102] To verify the effectiveness of this system, 30 elderly subjects were selected and divided into an experimental group (using the system of this invention) and a control group (using a traditional smart pillbox + a wristband with fixed 100Hz PPG sampling). Monitoring was conducted continuously for 14 days, and the results are as follows: Analysis of experimental results: As shown in the simulation diagram of multimodal medication action perception, traditional smart pillboxes can only monitor the decrease in tray weight and cannot confirm the final destination of the medication, resulting in a deception prevention rate of only 32.5% in the control group. This invention employs a multimodal perception mechanism: such as... Figure 3 As shown in (a), the system first detects a load step ( ); then as Figure 3 As shown in (b), the system compares the extreme values ​​of the cross-correlation of spatial motion trajectories to capture the contact motion. ); Finally, as Figure 3 As shown in (c), the extreme values ​​of the dual-mode cross-energy of mechanical vibration and sound waves are extracted to confirm the swallowing action. The three action nodes must satisfy the set physical causal time difference (i.e., Furthermore, the intervals are within the range of physiological constants. Through a rigorous temporal coherence logic chain, the anti-fraud recognition rate of medication behavior in the experimental group reached 98.3%, solving the management pain point of special groups (such as elderly people with cognitive impairment, mental illness patients who refuse to take medication) not taking their medication or hiding their medication.

[0103] like Figure 4 As shown, PPG physiological pulse waves require a high sampling rate, but continuous high-frequency sampling can deplete the wearable device's battery. This solution breaks away from the traditional fixed polling mechanism, utilizing the real-time extrapolated first-order time derivative of blood drug concentration (i.e., the baseline rate of change) as the control variable for the underlying hardware registers. In the figure, when the expected blood drug concentration (solid line) is in the stage of rapid blood entry causing drastic changes in vascular resistance, the system controls the underlying circuitry to instantly increase the sampling frequency (dashed line) to capture the extreme values ​​of morphological gradients with high resolution; during the stable period of drug effect, the sampling rate drops back to the baseline. With this mechanism, the average daily power consumption of the wrist device in the experimental group dropped sharply from 450mAh to 110mAh, improving battery life by approximately 309%, without losing any key physiological features.

[0104] like Figure 5 As shown, in situations where delayed medication is frequently encountered by the elderly, traditional timed pillboxes lack metabolic tracking. The mechanical, timed opening can easily lead to an accumulation of blood drug concentrations from two separate doses, exceeding safe limits (as shown by the dotted line in the figure, resulting in pathological drug effect superposition). This invention extracts the actual physical swallowing time and incorporates continuously adaptive and iterative individualized... A constant is used to accurately calculate the residual load in the body. When an overdose is predicted for the next administration, the system not only issues a software alarm but also directly pulls down the pin level via the microcontroller to rigidly electromechanically lock the dispensing compartment door that has not yet been opened. Figure 5 As shown by the solid line, the system proactively postponed the originally scheduled activation command until an absolutely safe time point. During the experimental period, the system successfully conducted 47 physical-level interception and early warnings, fundamentally ensuring the safety of drug users.

Claims

1. An intelligent medication management system based on multimodal perception and adaptive decision-making, characterized in that, include: An automated dispensing terminal has a weighing sensor, a dispensing chamber, and an electromechanical locking mechanism for controlling the opening and closing of the dispensing chamber; The wrist-mounted edge computing device is communicatively connected to the automated drug dispensing terminal and integrates an inertial measurement unit, an acoustic pickup module, and an optical hemodynamic sensing module. The wrist edge computing device establishes a feature vector based on the load step time obtained by the weighing sensor, and the contact action time and swallowing action time extracted by the inertial measurement unit and the acoustic pickup module, and records the preliminary medication confirmation status when the set conditions are met. Based on the swallowing action time, the pharmacokinetic model is retrieved to extrapolate the expected concentration change rate, and the sampling frequency of the optical hemodynamic sensing module is dynamically adjusted to collect physiological signals. The closed-loop verification pass status is determined based on the morphological gradient sequence generated by the physiological signal, and the electromechanical locking mechanism is controlled accordingly.

2. The intelligent medication management system based on multimodal perception and adaptive decision-making according to claim 1, characterized in that, The extraction contact action time and swallowing action time include: The spatial attitude sequence acquired by the inertial measurement unit is cross-correlated with the standard motion trajectory sequence, and the time corresponding to the extreme value is extracted as the contact action time. After the contact action time, acoustic wave signals and mechanical vibration signals are collected and cross-energy functions are calculated. The time corresponding to the extreme value is extracted as the swallowing action time.

3. The intelligent medication management system based on multimodal perception and adaptive decision-making according to claim 1, characterized in that, The projected expected concentration change rate and the adjustment of the sampling frequency include: Based on the swallowing action time as the starting point, the pharmacokinetic model is used to construct a blood drug concentration function and calculate the time derivative to obtain concentration change rate data. The absolute value of the concentration change rate data is converted into a sampling frequency command using a mapping function, thereby controlling the actual sampling frequency of the optical hemodynamic sensing module.

4. The intelligent medication management system based on multimodal perception and adaptive decision-making according to claim 1, characterized in that, The determination of the closed-loop verification pass status includes: Calculate the extreme value difference of the morphological gradient sequence. When the preliminary medication confirmation state exists and the extreme value difference is greater than or equal to the basic physiological fluctuation tolerance, record the closed-loop verification pass status. The actual physical peak time of the morphological gradient sequence is extracted, the time difference between the actual physical peak time and the theoretical concentration peak time is calculated, and the absorption rate constant in the pharmacokinetic model is corrected based on the time difference.

5. The intelligent medication management system based on multimodal perception and adaptive decision-making according to claim 4, characterized in that, The control mechanism for the electromechanical locking includes: When the difference between the extreme values ​​is less than the basic physiological fluctuation tolerance, the electromechanical locking mechanism is controlled to lock the unopened dispensing compartment. When the swallowing action time is later than the planned dosing time, the residual drug load is calculated using the corrected absorption rate constant and then superimposed with the next expected blood drug concentration to generate the estimated superimposed peak. When the estimated superimposed peak value is greater than or equal to the upper limit of safe concentration, the start time of the next simulation is shifted forward on the time axis until the estimated superimposed peak value is less than the upper limit of safe concentration, the next safe dosing time node is obtained, and the electromechanical locking mechanism is controlled to lock to the next safe dosing time node.

6. An intelligent medication management method based on multimodal perception and adaptive decision-making, characterized in that, The intelligent medication management system based on multimodal perception and adaptive decision-making, applied to any one of claims 1-5, comprises: The load step time of the automated dispensing terminal, as well as the contact action time and swallowing action time extracted by the wrist device, are obtained. A feature vector is established and the initial medication confirmation status is recorded when the conditions are met. Based on the swallowing action time, the pharmacokinetic model is retrieved to extrapolate the expected concentration change rate, and the sampling frequency of the optical sensing module is dynamically adjusted to collect physiological signals. The closed-loop verification status is determined based on the morphological gradient sequence generated by the physiological signal, and the electromechanical locking mechanism of the drug dispensing terminal is controlled.

7. The intelligent medication management method based on multimodal perception and adaptive decision-making according to claim 6, characterized in that, The acquisition of the contact action time and swallowing action time includes: Cross-correlation is performed between the spatial attitude sequence and the standard motion trajectory sequence, and the time corresponding to the extreme value is extracted as the contact action time. After the contact action time, the acoustic signal and mechanical vibration signal are collected synchronously and the cross energy function is calculated. The time corresponding to the extreme value is extracted as the swallowing action time.

8. The intelligent medication management method based on multimodal perception and adaptive decision-making according to claim 6, characterized in that, The dynamic adjustment of the sampling frequency includes: Based on the swallowing action time as the starting point, the concentration time derivative is calculated using the pharmacokinetic model to generate concentration change rate data. The absolute value of the concentration change rate data is mapped to a sampling frequency command to control the actual sampling frequency.

9. The intelligent medication management method based on multimodal perception and adaptive decision-making according to claim 6, characterized in that, The determination of the closed-loop verification pass status includes: Calculate the difference between the extreme values ​​of the morphological gradient sequence. When the difference between the extreme values ​​is greater than or equal to the basic physiological fluctuation tolerance, record the closed-loop verification status. Extract the actual physical peak time, calculate the time difference between the actual physical peak time and the theoretical concentration peak time, and correct the absorption rate constant in the pharmacokinetic model based on the time difference.

10. The intelligent medication management method based on multimodal perception and adaptive decision-making according to claim 9, characterized in that, The control electromechanical locking mechanism includes: When the difference between the extreme values ​​is less than the basic physiological fluctuation tolerance, the electromechanical locking mechanism is controlled to lock the dispensing chamber. When delayed medication occurs, the residual drug load is calculated using the corrected parameters and the superimposed peak value is estimated. When the estimated peak value exceeds the upper limit of the safe concentration, the start time of the next administration is shifted forward until the estimated peak value is less than the upper limit of the safe concentration. The time node for the next safe administration is determined and the electromechanical locking mechanism is controlled to lock to the next safe administration time node.

Images